Researchers introduce the Foveated Dynamic Transformer (FDT), a vision transformer architecture inspired by the human visual system's foveated sampling and eye movements. The model integrates fixation and foveation modules to dynamically select tokens, filtering irrelevant information while generating multi-scale embeddings.
- FDT achieves 81.9% accuracy compared to DeiT-S's 80.9% at a 50% fixation-budget setting.
- The architecture reduces multiply-accumulate operations by 34.57% while maintaining higher accuracy.
- The model exhibits strong resilience to various types of noise and adversarial attacks without explicit training for these challenges.
These attributes position FDT as a step toward artificial neural networks that combine adaptive computation with improved resilience.